Large language models (LLMs) can sometimes report the strategies they actually use to solve tasks, yet at other times seem unable to recognize those strategies that govern their behavior. This suggests a limited degree of metacognition --- the capacity to monitor one's own cognitive processes for subsequent reporting and self-control. Metacognition enhances LLMs' capabilities in solving complex tasks but also raises safety concerns, as models may obfuscate their internal processes to evade neural-activation-based oversight (e.g., safety detector). Given society's increased reliance on these models, it is critical that we understand their metacognitive abilities. To address this, we introduce a neuroscience-inspired \emph{neurofeedback} paradigm that uses in-context learning to quantify metacognitive abilities of LLMs to \textit{report} and \textit{control} their activation patterns. We demonstrate that their abilities depend on several factors: the number of in-context examples provided, the semantic interpretability of the neural activation direction (to be reported/controlled), and the variance explained by that direction. These directions span a ``metacognitive space'' with dimensionality much lower than the model's neural space, suggesting LLMs can monitor only a small subset of their neural activations. Our paradigm provides empirical evidence to quantify metacognition in LLMs, with significant implications for AI safety (e.g., adversarial attack and defense).

Self-play fine-tuning has demonstrated promising abilities in adapting large language models (LLMs) to downstream tasks with limited real-world data. The basic principle is to iteratively refine the model with real samples and synthetic ones generated from itself. However, the existing methods primarily focus on the relative gaps between the rewards for two types of data, neglecting their absolute values. Through theoretical analysis, we identify that the gap-based methods suffer from unstable evolution, due to the potentially degenerated objectives.

Recent endeavors in Multimodal Large Language Models (MLLMs) aim to unify visual comprehension and generation. However, these two capabilities remain largely independent, as if they are two separate functions encapsulated within the same model. Consequently, visual comprehension does not enhance visual generation, and the reasoning mechanisms of LLMs have not been fully integrated to revolutionize image generation. In this paper, we propose to enable the collaborative co-evolution of visual comprehension and generation, advancing image generation into an iterative introspective process. We introduce a two-stage training approach: supervised fine-tuning teaches the MLLM with the foundational ability to generate genuine CoT for visual generation, while reinforcement learning activates its full potential via an exploration-exploitation trade-off. Ultimately, we unlock the Aha moment in visual generation, advancing MLLMs from text-to-image tasks to unified image generation. Extensive experiments demonstrate that our model not only excels in text-to-image generation and image editing, but also functions as a superior image semantic evaluator with enhanced visual comprehension capabilities.

Multimodal large language models (MLLMs) have achieved remarkable progress in video understanding. However, hallucination, where the model generates plausible yet incorrect outputs, persists as a significant and under-addressed challenge in the video domain. Among existing solutions, activation engineering has proven successful in mitigating hallucinations in LLMs and ImageLLMs, yet its applicability to VideoLLMs remains largely unexplored. In this work, we are the first to systematically investigate the effectiveness and underlying mechanisms of activation engineering for mitigating hallucinations in VideoLLMs. We initially conduct an investigation of the key factors affecting the performance of activation engineering and find that a model's sensitivity to hallucination depends on $\textbf{temporal variation}$ rather than task type. Moreover, selecting appropriate internal modules and dataset for activation engineering is critical for reducing hallucination. Guided by these findings, we propose a temporal-aware activation engineering framework for VideoLLMs, which adaptively identifies and manipulates hallucination-sensitive modules based on the temporal variation characteristic, substantially mitigating hallucinations without additional LLM fine-tuning. Experiments across multiple models and benchmarks demonstrate that our method markedly reduces hallucination in VideoLLMs, thereby validating the robustness of our findings.

Constrained decoding enables Language Models (LMs) to produce samples that provably satisfy hard constraints. However, existing constrained-decoding approaches often distort the underlying model distribution, a limitation that is especially problematic in applications like program fuzzing, where one wants to generate diverse and valid program inputs for testing purposes. We propose a new constrained sampling framework based on Markov Chain Monte Carlo (MCMC) that simultaneously satisfies three core desiderata: constraint satisfying (every sample satisfies the constraint), monotonically converging (the sampling process converges to the true conditional distribution), and efficient (high-quality samples emerge in few steps). Our method constructs a proposal distribution over valid outputs and applies a Metropolis-Hastings acceptance criterion based on the LM's likelihood, ensuring principled and efficient exploration of the constrained space. Empirically, our sampler outperforms existing methods on both synthetic benchmarks and real-world program fuzzing tasks.

Finding low-rank matrix weights is a key technique for addressing the high memory usage and computational demands of large models. Most existing algorithms rely on the factorization of the low-rank matrix weights, which is non-unique and redundant. Their convergence is slow especially when the target low-rank matrices are ill-conditioned, because the convergence rate depends on the condition number of the Jacobian operator for the factorization and the Hessian of the loss function with respect to the weight matrix. To address this challenge, we adopt the Riemannian gradient descent (RGD) algorithm on the Riemannian manifold of fixed-rank matrices to update the entire low-rank weight matrix. This algorithm completely avoids the factorization, thereby eliminating the negative impact of the Jacobian condition number.

This paper introduces an approach for training o1-like RAG models that retrieve and reason over relevant information step by step before generating the final answer. Conventional RAG methods usually perform a single retrieval step before the generation process, which limits their effectiveness in addressing complex queries due to imperfect retrieval results. In contrast, our proposed method, CoRAG (Chain-of-Retrieval Augmented Generation), allows the model to dynamically reformulate the query based on the evolving state. To train CoRAG effectively, we utilize rejection sampling to automatically generate intermediate retrieval chains, thereby augmenting existing RAG datasets that only provide the correct final answer. At test time, we propose various decoding strategies to scale the model's test-time compute by controlling the length and number of sampled retrieval chains. Experimental results across multiple benchmarks validate the efficacy of CoRAG, particularly in multi-hop question answering tasks, where we observe more than $10$ points improvement in EM score compared to strong baselines. On the KILT benchmark, CoRAG establishes a new state-of-the-art performance across a diverse range of knowledge-intensive tasks. Furthermore, we offer comprehensive analyses to understand the scaling behavior of CoRAG, laying the groundwork for future research aimed at developing factual and grounded foundation models.

Recent advances of Reinforcement Learning (RL) have highlighted its potential in complex reasoning tasks, yet effective training often relies on external supervision, which limits the broader applicability. In this work, we propose a novel self-rewarding reinforcement learning framework to enhance Large Language Model (LLM) reasoning by leveraging the consistency of intermediate reasoning states across different reasoning trajectories. Our key insight is that correct responses often exhibit consistent trajectory patterns in terms of model likelihood: their intermediate reasoning states tend to converge toward their own final answers () with minimal deviation toward other candidates ().

We introduce BayeSQP, a novel algorithm for general black-box optimization that merges the structure of sequential quadratic programming with concepts from Bayesian optimization. BayeSQP employs second-order Gaussian process surrogates for both the objective and constraints to jointly model the function values, gradients, and Hessian from only zero-order information. At each iteration, a local subproblem is constructed using the GP posterior estimates and solved to obtain a search direction. Crucially, the formulation of the subproblem explicitly incorporates uncertainty in both the function and derivative estimates, resulting in a tractable second-order cone program for high probability improvements under model uncertainty. A subsequent one-dimensional line search via constrained Thompson sampling selects the next evaluation point. Empirical results show that BayeSQP outperforms state-of-the-art methods in specific high-dimensional settings. Our algorithm offers a principled and flexible framework that bridges classical optimization techniques with modern approaches to black-box optimization.

Diffusion models have recently gained prominence in offline reinforcement learning due to their ability to effectively learn high-performing, generalizable policies from static datasets. Diffusion-based planners facilitate long-horizon decision-making by generating high-quality trajectories through iterative denoising, guided by return-maximizing objectives. However, existing guided sampling strategies such as Classifier Guidance, Classifier-Free Guidance, and Monte Carlo Sample Selection either produce suboptimal multi-modal actions, struggle with distributional drift, or incur prohibitive inference-time costs. To address these challenges, we propose \textbf{\textit{Prior Guidance}} (PG), a novel guided sampling framework that replaces the standard Gaussian prior of a behavior-cloned diffusion model with a learnable distribution, optimized via a behavior-regularized objective. PG directly generates high-value trajectories without costly reward optimization of the diffusion model itself, and eliminates the need to sample multiple candidates at inference for sample selection. We present an efficient training strategy that applies behavior regularization in latent space, and empirically demonstrate that PG outperforms state-of-the-art diffusion policies and planners across diverse long-horizon offline RL benchmarks. Our code is available at https://github.com/ku-dmlab/PG.